File size: 13,781 Bytes
b887ad8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
import sys
from datetime import datetime
import numpy as np
import torch
from utils.metrics import *
from collections import OrderedDict
import os



def evaluate_matching_score(eval_wrapper,motion_loaders, file):
    match_score_dict = OrderedDict({})
    R_precision_dict = OrderedDict({})
    activation_dict = OrderedDict({})
    # print(motion_loaders.keys())
    print('========== Evaluating Matching Score ==========')
    for motion_loader_name, motion_loader in motion_loaders.items():
        all_motion_embeddings = []
        score_list = []
        all_size = 0
        matching_score_sum = 0
        top_k_count = 0
        # print(motion_loader_name)
        with torch.no_grad():
            for idx, batch in enumerate(motion_loader):
                word_embeddings, pos_one_hots, _, sent_lens, motions, m_lens, _ = batch
                text_embeddings, motion_embeddings = eval_wrapper.get_co_embeddings(
                    word_embs=word_embeddings,
                    pos_ohot=pos_one_hots,
                    cap_lens=sent_lens,
                    motions=motions,
                    m_lens=m_lens
                )
                dist_mat = euclidean_distance_matrix(text_embeddings.cpu().numpy(),
                                                     motion_embeddings.cpu().numpy())
                matching_score_sum += dist_mat.trace()
                # import pdb;pdb.set_trace()

                argsmax = np.argsort(dist_mat, axis=1)
                top_k_mat = calculate_top_k(argsmax, top_k=3)
                top_k_count += top_k_mat.sum(axis=0)

                all_size += text_embeddings.shape[0]

                all_motion_embeddings.append(motion_embeddings.cpu().numpy())

            all_motion_embeddings = np.concatenate(all_motion_embeddings, axis=0)
            # import pdb;pdb.set_trace()
            matching_score = matching_score_sum / all_size
            R_precision = top_k_count / all_size
            match_score_dict[motion_loader_name] = matching_score
            R_precision_dict[motion_loader_name] = R_precision
            activation_dict[motion_loader_name] = all_motion_embeddings

        print(f'---> [{motion_loader_name}] Matching Score: {matching_score:.4f}')
        print(f'---> [{motion_loader_name}] Matching Score: {matching_score:.4f}', file=file, flush=True)

        line = f'---> [{motion_loader_name}] R_precision: '
        for i in range(len(R_precision)):
            line += '(top %d): %.4f ' % (i+1, R_precision[i])
        print(line)
        print(line, file=file, flush=True)

    return match_score_dict, R_precision_dict, activation_dict


def evaluate_fid(eval_wrapper,groundtruth_loader, activation_dict, file):
    eval_dict = OrderedDict({})
    gt_motion_embeddings = []
    print('========== Evaluating FID ==========')
    with torch.no_grad():
        for idx, batch in enumerate(groundtruth_loader):
            _, _, _, sent_lens, motions, m_lens, _ = batch
            motion_embeddings = eval_wrapper.get_motion_embeddings(
                motions=motions,
                m_lens=m_lens
            )
            gt_motion_embeddings.append(motion_embeddings.cpu().numpy())
    gt_motion_embeddings = np.concatenate(gt_motion_embeddings, axis=0)
    gt_mu, gt_cov = calculate_activation_statistics(gt_motion_embeddings)

    for model_name, motion_embeddings in activation_dict.items():
        mu, cov = calculate_activation_statistics(motion_embeddings)
        # print(mu)
        fid = calculate_frechet_distance(gt_mu, gt_cov, mu, cov)
        print(f'---> [{model_name}] FID: {fid:.4f}')
        print(f'---> [{model_name}] FID: {fid:.4f}', file=file, flush=True)
        eval_dict[model_name] = fid
    return eval_dict


def evaluate_diversity(activation_dict, file, diversity_times):
    eval_dict = OrderedDict({})
    print('========== Evaluating Diversity ==========')
    for model_name, motion_embeddings in activation_dict.items():
        diversity = calculate_diversity(motion_embeddings, diversity_times)
        eval_dict[model_name] = diversity
        print(f'---> [{model_name}] Diversity: {diversity:.4f}')
        print(f'---> [{model_name}] Diversity: {diversity:.4f}', file=file, flush=True)
    return eval_dict


def evaluate_multimodality(eval_wrapper, mm_motion_loaders, file, mm_num_times):
    eval_dict = OrderedDict({})
    print('========== Evaluating MultiModality ==========')
    for model_name, mm_motion_loader in mm_motion_loaders.items():
        mm_motion_embeddings = []
        with torch.no_grad():
            for idx, batch in enumerate(mm_motion_loader):
                # (1, mm_replications, dim_pos)
                motions, m_lens = batch
                motion_embedings = eval_wrapper.get_motion_embeddings(motions[0], m_lens[0])
                mm_motion_embeddings.append(motion_embedings.unsqueeze(0))
        if len(mm_motion_embeddings) == 0:
            multimodality = 0
        else:
            mm_motion_embeddings = torch.cat(mm_motion_embeddings, dim=0).cpu().numpy()
            multimodality = calculate_multimodality(mm_motion_embeddings, mm_num_times)
        print(f'---> [{model_name}] Multimodality: {multimodality:.4f}')
        print(f'---> [{model_name}] Multimodality: {multimodality:.4f}', file=file, flush=True)
        eval_dict[model_name] = multimodality
    return eval_dict


def get_metric_statistics(values, replication_times):
    mean = np.mean(values, axis=0)
    std = np.std(values, axis=0)
    conf_interval = 1.96 * std / np.sqrt(replication_times)
    return mean, conf_interval


def evaluation(eval_wrapper, gt_loader, eval_motion_loaders, log_file, replication_times, diversity_times, mm_num_times, run_mm=False):
    with open(log_file, 'a') as f:
        all_metrics = OrderedDict({'Matching Score': OrderedDict({}),
                                   'R_precision': OrderedDict({}),
                                   'FID': OrderedDict({}),
                                   'Diversity': OrderedDict({}),
                                   'MultiModality': OrderedDict({})})

        for replication in range(replication_times):
            print(f'Time: {datetime.now()}')
            print(f'Time: {datetime.now()}', file=f, flush=True)
            motion_loaders = {}
            motion_loaders['ground truth'] = gt_loader
            mm_motion_loaders = {}
            # motion_loaders['ground truth'] = gt_loader
            for motion_loader_name, motion_loader_getter in eval_motion_loaders.items():
                motion_loader, mm_motion_loader,eval_generate_time = motion_loader_getter()
                print(f'---> [{motion_loader_name}] batch_generate_time: {eval_generate_time}s', file=f, flush=True)
                motion_loaders[motion_loader_name] = motion_loader
                mm_motion_loaders[motion_loader_name] = mm_motion_loader

            if replication_times>1:
                print(f'==================== Replication {replication} ====================')
                print(f'==================== Replication {replication} ====================', file=f, flush=True)


            mat_score_dict, R_precision_dict, acti_dict = evaluate_matching_score(eval_wrapper, motion_loaders, f)

            fid_score_dict = evaluate_fid(eval_wrapper, gt_loader, acti_dict, f)

            div_score_dict = evaluate_diversity(acti_dict, f, diversity_times)

            if run_mm:
                mm_score_dict = evaluate_multimodality(eval_wrapper, mm_motion_loaders, f, mm_num_times)

            print(f'!!! DONE !!!')
            print(f'!!! DONE !!!', file=f, flush=True)

            for key, item in mat_score_dict.items():
                if key not in all_metrics['Matching Score']:
                    all_metrics['Matching Score'][key] = [item]
                else:
                    all_metrics['Matching Score'][key] += [item]

            for key, item in R_precision_dict.items():
                if key not in all_metrics['R_precision']:
                    all_metrics['R_precision'][key] = [item]
                else:
                    all_metrics['R_precision'][key] += [item]

            for key, item in fid_score_dict.items():
                if key not in all_metrics['FID']:
                    all_metrics['FID'][key] = [item]
                else:
                    all_metrics['FID'][key] += [item]

            for key, item in div_score_dict.items():
                if key not in all_metrics['Diversity']:
                    all_metrics['Diversity'][key] = [item]
                else:
                    all_metrics['Diversity'][key] += [item]

            for key, item in mm_score_dict.items():
                if key not in all_metrics['MultiModality']:
                    all_metrics['MultiModality'][key] = [item]
                else:
                    all_metrics['MultiModality'][key] += [item]


        mean_dict = {}
        if replication_times>1:
            for metric_name, metric_dict in all_metrics.items():
                print('========== %s Summary ==========' % metric_name)
                print('========== %s Summary ==========' % metric_name, file=f, flush=True)

                for model_name, values in metric_dict.items():
                    # print(metric_name, model_name)
                    mean, conf_interval = get_metric_statistics(np.array(values),replication_times)
                    mean_dict[metric_name + '_' + model_name] = mean
                    # print(mean, mean.dtype)
                    if isinstance(mean, np.float64) or isinstance(mean, np.float32):
                        print(f'---> [{model_name}] Mean: {mean:.4f} CInterval: {conf_interval:.4f}')
                        print(f'---> [{model_name}] Mean: {mean:.4f} CInterval: {conf_interval:.4f}', file=f, flush=True)
                    elif isinstance(mean, np.ndarray):
                        line = f'---> [{model_name}]'
                        for i in range(len(mean)):
                            line += '(top %d) Mean: %.4f CInt: %.4f;' % (i+1, mean[i], conf_interval[i])
                        print(line)
                        print(line, file=f, flush=True)
            return mean_dict
        else:
            return all_metrics


def distributed_evaluation(eval_wrapper, gt_loader, eval_motion_loader, log_file, replication_times, diversity_times):
    with open(log_file, 'a') as f:
        all_metrics = OrderedDict({'Matching Score': OrderedDict({}),
                                   'R_precision': OrderedDict({}),
                                   'FID': OrderedDict({}),
                                   'Diversity': OrderedDict({}),
                                   'MultiModality': OrderedDict({})})

        for replication in range(replication_times):
            print(f'Time: {datetime.now()}')
            print(f'Time: {datetime.now()}', file=f, flush=True)
            motion_loaders = {'test':eval_motion_loader}

            if replication_times>1:
                print(f'==================== Replication {replication} ====================')
                print(f'==================== Replication {replication} ====================', file=f, flush=True)


            mat_score_dict, R_precision_dict, acti_dict = evaluate_matching_score(eval_wrapper, motion_loaders, f)

            fid_score_dict = evaluate_fid(eval_wrapper, gt_loader, acti_dict, f)

            div_score_dict = evaluate_diversity(acti_dict, f, diversity_times)


            print(f'!!! DONE !!!')
            print(f'!!! DONE !!!', file=f, flush=True)

            for key, item in mat_score_dict.items():
                if key not in all_metrics['Matching Score']:
                    all_metrics['Matching Score'][key] = [item]
                else:
                    all_metrics['Matching Score'][key] += [item]

            for key, item in R_precision_dict.items():
                if key not in all_metrics['R_precision']:
                    all_metrics['R_precision'][key] = [item]
                else:
                    all_metrics['R_precision'][key] += [item]

            for key, item in fid_score_dict.items():
                if key not in all_metrics['FID']:
                    all_metrics['FID'][key] = [item]
                else:
                    all_metrics['FID'][key] += [item]

            for key, item in div_score_dict.items():
                if key not in all_metrics['Diversity']:
                    all_metrics['Diversity'][key] = [item]
                else:
                    all_metrics['Diversity'][key] += [item]

        mean_dict = {}
        for metric_name, metric_dict in all_metrics.items():
            print('========== %s Summary ==========' % metric_name)
            print('========== %s Summary ==========' % metric_name, file=f, flush=True)

            for model_name, values in metric_dict.items():
                # print(metric_name, model_name)
                mean, conf_interval = get_metric_statistics(np.array(values),replication_times)
                mean_dict[metric_name + '_' + model_name] = mean
                # print(mean, mean.dtype)
                if isinstance(mean, np.float64) or isinstance(mean, np.float32):
                    print(f'---> [{model_name}] Mean: {mean:.4f} CInterval: {conf_interval:.4f}')
                    print(f'---> [{model_name}] Mean: {mean:.4f} CInterval: {conf_interval:.4f}', file=f, flush=True)
                elif isinstance(mean, np.ndarray):
                    line = f'---> [{model_name}]'
                    for i in range(len(mean)):
                        line += '(top %d) Mean: %.4f CInt: %.4f;' % (i+1, mean[i], conf_interval[i])
                    print(line)
                    print(line, file=f, flush=True)
        return mean_dict